Transcoders Beat Sparse Autoencoders for Interpretability

TL;DR

This paper compares transcoders and sparse autoencoders in deep neural networks, demonstrating that transcoders produce more interpretable features and introducing skip transcoders that improve reconstruction without sacrificing interpretability.

Contribution

It introduces skip transcoders and provides a comparative analysis showing their superior interpretability over sparse autoencoders.

Findings

Transcoders yield more interpretable features than SAEs.

Skip transcoders achieve lower reconstruction loss.

Skip transcoders do not reduce interpretability.

Abstract

Sparse autoencoders (SAEs) extract human-interpretable features from deep neural networks by transforming their activations into a sparse, higher dimensional latent space, and then reconstructing the activations from these latents. Transcoders are similar to SAEs, but they are trained to reconstruct the output of a component of a deep network given its input. In this work, we compare the features found by transcoders and SAEs trained on the same model and data, finding that transcoder features are significantly more interpretable. We also propose skip transcoders, which add an affine skip connection to the transcoder architecture, and show that these achieve lower reconstruction loss with no effect on interpretability.